New Framework Enhances Multimodal Graph Learning with Context-Aware Alignment

Multimodal Attributed Graphs (MAGs) are used to model real-world entities by combining graph structures with diverse attributes like text and images. Current MAG methods often struggle with fixed graph contexts or over-compressed data fusion, which can limit their effectiveness for various tasks requiring both structural and fine-grained cross-modal understanding. A new framework called CoMAG has been introduced to address these challenges. CoMAG learns reliable contexts that adapt to specific tasks by estimating edge reliability from multimodal semantic consistency and selecting context components via a task-aware gate. It also performs modality-preserving alignment by maintaining distinct multi-hop trajectories for each modality and decoupling shared from private representations. This approach allows CoMAG to generate both graph and modality representations in a single pass, preserving modality-specific information. Experimental results across multiple datasets show CoMAG outperforms existing baselines in graph-level prediction, modality matching, and graph-conditioned generation, while maintaining efficient complexity.